Open YuhuiMa opened 3 years ago
In the code, loaded samples seem only to substract mean but not to be normalized, which leads to very large loss. Why not normalize the samples?
The VGG model I used was converted from Caffe, which was trained without std normalization.
In the code, loaded samples seem only to substract mean but not to be normalized, which leads to very large loss. Why not normalize the samples?